REVIEW 4 major objections 6 minor 2 references
Mixed pooling of seasonality for time series forecasting: An application to pallet transport data
T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read A mixed hierarchical seasonality model learns each seasonality's weight from data and forecasts pallet transport more accurately than complete pooling, Fourier decomposition, or SARIMA.
desk verdict The modeling idea is coherent, but the paper's central empirical claim fails on its own numbers: the mixed model loses to Fourier on the only multiple-seasonality dataset, and the ELPD comparison is sign-inverted. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The engine is the mixed pooling forecast, in which the slope and intercept at time $t$ are $\sum_{d=1}^D \theta_d\, k_{d,\,j_d(t)}$ and $\sum_{d=1}^D \theta_d\, m_{d,\,j_d(t)}$, where $d$ indexes the seasonal patterns (week, month), $j_d(t)$ is the subcategory (weekday or month-day) of time $t$ in pattern $d$, and the parameters $k$ and $m$ are drawn from shared normal hyperpriors. The hierarchical priors give partial pooling, so a distinctive period such as Sunday or month-end keeps its own shape while borrowing strength from other periods. The $\theta$ vector is a simplex, with components summing to one, and is learned from the data, so the model automatically sets the relative importance of each seasonality. This weighted-average construction, rather than an additive sum of seasonal effects, is what gives the model its name and its claimed ability to handle interactions.
What would settle it
A decisive check is already in the paper's Table 1: for the shipment flow, the mixed pooling MAPE (0.95) is worse than Fourier decomposition (0.90), so a replication on additional multi-seasonal datasets, with uncertainty intervals around MAPE, would settle whether the claimed superiority over Fourier is real; a complementary simulation with a multiplicative weekday-month interaction would test whether the linear-mixture assumption is identifiable.
Extended reading notes
Core claim
The central claim is that seasonal interactions can be learned rather than pre-specified. In the mixed pooling model, subcategory parameters $k_{ij}$ and $m_{ij}$ stand for the slope and intercept of day-of-week or day-of-month $j$ within seasonality $i$, sharing normal hyperpriors; the forecast mean is a weighted average of those per-period vectors with a simplex weight $\theta$. This is the mechanism that lets the model represent the combined effect of week and month as a learned mixture instead of an assumed sum. On the delivery and restocking flows, which have a single dominant season, the paper reports that partial pooling cuts MAPE sharply (delivery from 292.13 to 31.17; restocking from 12.66 to 4.09). On the shipment flow, where weekly and monthly rhythms interact, it reports that mixed pooling performed best and that the posterior weight, roughly $\theta_{\text{week}} = 0.6$ and $\theta_{\text{month}} = 0.4$, identifies weekly seasonality as the larger contributor.
Load-bearing premise
The model treats the interaction between weekly and monthly seasonality as a single weighted average of separately estimated trend lines, with one global weight per season applied to both slope and intercept, so if the true interaction is not that simple linear mixture, the forecasts are structurally wrong; the paper tests this on only one shipment dataset.
Editorial extensions
If this is right
- For single-seasonality series, partial pooling alone captures periodic scale jumps; the delivery MAPE drops from 292.13 (complete pooling) to 31.17 (week pooling).
- For multi-seasonality series, forecasters no longer need to guess which season dominates: the model learns the relative strength from data, unlike SARIMA or Fourier setups that require the seasonal form to be fixed in advance.
- The learned $\theta$ gives an operational handle: managers can rank weekly versus monthly effects and time replenishment or staffing decisions accordingly.
- Because the model only assumes an interchangeable hierarchy, the same construction extends to hourly, daily, quarterly, or yearly seasonalities by adding dimensions to the parameter arrays and the simplex.
Reading between the lines
- The per-period posterior distributions double as a diagnostic for anomalous calendar days; for instance, the paper's own parameter plots single out Sunday in delivery and month boundaries in restocking, which could be used to flag calendar effects the model cannot explain.
- If the linear mixture is correct, then the mixture weight $\theta$ should be stable across rolling windows for a stationary seasonal regime; a drifting $\theta$ would be a simple, testable early warning for changing seasonal behavior.
- A natural extension is to treat $\theta$ as a group-level parameter shared across related series, such as all flows in a logistics network, shrinking individual weights toward a common seasonal profile and borrowing strength when one flow has little data.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a Bayesian 'mixed hierarchical seasonality' (MHS) model for time series with multiple seasonal patterns. The model gives each day-of-week and day-of-month subcategory its own trend slope and intercept, partially pools those parameters through hyperpriors, and combines the weekly and monthly seasonal components via a simplex weight vector theta. The authors compare partial pooling (for single-seasonality series) and mixed pooling (for the series with both weekly and monthly seasonality) against complete pooling, Fourier decomposition, and SARIMA on three daily pallet-transport series from a logistics network. They report MAPE and ELPD as out-of-sample measures and claim that the proposed models achieve 'considerable improvements' over the baselines, especially for the shipment series, which exhibits both weekly and monthly seasonality.
Significance. If the reported improvements were reliable, the MHS model would offer a useful and interpretable alternative for business time series with strong, multiple seasonal cycles. The paper includes Stan code in Appendix A and applies the model to a real logistics dataset, which are positive features. However, the significance is seriously undermined by the fact that the paper's own tables contradict its central empirical claims: the mixed pooling model loses to Fourier decomposition on the one dataset where it is used, and the ELPD table appears to be interpreted with an inverted sign convention. As presented, the paper does not establish the claimed advantage of the proposed approach.
major comments (4)
- [Section 5.1, Table 1] For the shipment dataset, the only dataset for which the mixed pooling model is used, the reported MAPE for mixed pooling is 0.95, while the Fourier decomposition model achieves 0.90. The text states that 'For data with multiple seasonality, shipment, our proposed mixed pooling model showed better performance,' which is directly contradicted by the table. Since the abstract's central claim of 'considerable improvements' over Fourier decomposition rests on this comparison, the main empirical result is not established.
- [Section 5.1, Table 2] The text states that 'Lower ELPD implies a better model,' but the loo package defines ELPD such that higher values indicate better out-of-sample predictive density. Under the correct sign convention, the table contradicts the paper's conclusions: for shipment, Fourier (-1053) beats mixed pooling (-1279); for restocking, Fourier (-962) beats partial pooling (-1409); and for delivery, complete pooling (268) beats both Fourier (-728) and partial pooling (-729). Thus the statement that 'Partial pooling and mixed pooling outperforms the other models for every dataset' is only true under the inverted sign convention and is false when ELPD is interpreted correctly.
- [Section 4, Section 5.1] The evaluation is based on a single proprietary dataset, and no data or code are released, so the results are not reproducible. Furthermore, the MAPE differences are reported without any uncertainty quantification across the 12 sliding-window test sets; for example, the delivery comparison (31.17 vs 31.33) and the shipment comparison (0.95 vs 0.90) could easily be within sampling noise. The choice of pooling structure (partial vs mixed) is made after inspecting the same data, which introduces a selection effect that is not accounted for in the reported comparisons.
- [Section 3.3, Eq. (4c)] The mixed pooling model assumes that the combined weekly and monthly seasonal effect is a linear mixture of separately estimated slope and intercept parameters with a single global weight theta. This linearity assumption is neither justified nor tested, and the shipment dataset—the only place where this assumption is exercised—shows the model underperforming the Fourier baseline. This outcome is consistent with the possibility that the interaction between the two seasonal patterns is not a simple weighted average, which is a fundamental misspecification concern for the proposed model.
minor comments (6)
- [Section 3.2] Equation (3b) appears twice; the second displayed equation should be numbered (3c) or later equations should be renumbered.
- [Section 4] The sentence 'The former three correspond to' is incomplete; the SARIMA parameter sets are listed but the clause is not finished.
- [Section 2.1] The word 'Mulitseasonal' should be 'multiseasonal'.
- [Figure 4 caption] The caption states that 'The mixed pooling model shows a significantly higher performance than the other models,' but Figure 4 displays restocking data, which was analyzed with partial pooling, not mixed pooling; this is misleading.
- [Table 1] Several cells contain stray periods (e.g., '31.17 (week pool) .'), which appear to be formatting artifacts and should be removed.
- [Introduction] The citation 'Hyndman, 2018' in the introduction appears to refer to Hyndman and Athanasopoulos (2018), but the reference list contains only the latter; the in-text citation should be corrected.
Circularity Check
No circularity: the MHS model is an empirical proposal evaluated by out-of-sample comparisons; the derivation chain is self-contained.
full rationale
The paper makes no claim to derive MHS from first principles; it proposes a hierarchical linear-mixture model (Eq. 4c) and evaluates it on sliding-window out-of-sample data. The mixture weight theta is fitted on training folds and used for forecasts, which is ordinary parameter estimation, not a fitted quantity renamed as a prediction. The MAPE and ELPD comparisons are external benchmarks, so the central empirical claim is testable rather than forced by construction. The only self-citation (Moon and Song 2019) is a literature-review reference to a prior clustering study and is not load-bearing for the MHS derivation; it does not supply a premise on which the results depend. The limitations section explicitly concedes single-domain validation and trend-change limitations, which are generalizability caveats rather than circular reductions. The serious problems with the paper are correctness and validity issues, not circularity: Table 1 shows mixed pooling (0.95 MAPE) does not beat Fourier (0.90) on the sole multiple-seasonality dataset, contradicting the abstract's superiority claim, and Table 2's statement 'Lower ELPD implies a better model' reverses the conventional ELPD orientation. Those issues belong to a correctness assessment, not to a finding of circular derivation.
Assumptions & free parameters
free parameters (2)
- theta =
simplex weight; posterior mode approx 0.6 weekly, 0.4 monthly for shipment (Fig. 7)
- hyperparameters k_sigma, m_sigma, sigma_obs =
not reported numerically
assumptions (3)
- ad hoc to paper Combined seasonal effect equals a linear mixture of separate weekday and month-day parameters (Eq. 4c).
- domain assumption Daily series follow a normal distribution with constant variance around a linear trend for each seasonal subperiod.
- domain assumption Weekday (7) and month-day (31) groupings capture the relevant seasonal structure.
Cite this review
Pith. "Pith review of Mixed pooling of seasonality for time series forecasting: An application to pallet transport data." pith.science (2026). https://pith.science/paper/V5KKWPI4
@misc{pith2026190805339,
author = {Pith},
title = {Pith review of: Mixed pooling of seasonality for time series forecasting: An application to pallet transport data},
year = {2026},
howpublished = {\url{https://pith.science/paper/V5KKWPI4}},
note = {Machine review of arXiv:1908.05339}
}
read the original abstract
Multiple seasonal patterns play a key role in time series forecasting, especially for business time series where seasonal effects are often dramatic. Previous approaches including Fourier decomposition, exponential smoothing, and seasonal autoregressive integrated moving average (SARIMA) models do not reflect the distinct characteristics of each period in seasonal patterns. We propose a mixed hierarchical seasonality (MHS) model. Intermediate parameters for each seasonal period are first estimated, and a mixture of intermediate parameters is taken. This results in a model that automatically learns the relative importance of each seasonality and addresses the interactions between them. The model is implemented with Stan, a probabilistic language, and was compared with three existing models on a real-world dataset of pallet transport from a logistic network. Our new model achieved considerable improvements in terms of out of sample prediction error (MAPE) and predictive density (ELPD) compared to complete pooling, Fourier decomposition, and SARIMA model.
Figures
Reference graph
Works this paper leans on
-
[1]
Partial pooling data { int<lower=1> T; int<lower=1> P; vector[T] t; vector[T] y; int pool[T]; } parameters { real k_mu; real<lower=0> k_sigma; real m_mu; 22 real<lower=0> m_sigma; real k[P]; real m[P]; real<lower=0> sigma_obs; } transformed parameters { vector[T] p_k; vector[T] p_m; vector[T] yhat; for (i in 1:T) { p_k[i] = k[pool[i]]; p_m[i] = m[pool[i]]...
-
[2]
Mixed pooling model (MHS) data { int<lower=1> T; int<lower=1> D; int<lower=1> P; vector[T] t; 23 vector[T] y; int pool[T, D]; } parameters { real k_mu; real<lower=0> k_sigma; real m_mu; real<lower=0> m_sigma; real k[D, P]; real m[D, P]; simplex[D] theta; real<lower=0> sigma_obs; } transformed parameters { matrix[T, D] p_k; matrix[T, D] p_m; vector[T] yhat...
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.